AI Report Library

Find a prompt, model or category.

Business use casesllm-business-practices

Real business uses of LLMs by non-engineers

Sonnet 5.5 Anthropic · Medium

As of

Content density chart
PromptReal business uses of LLMs by non-engineers

list the practical real business examples by non-engineering people with LLM.

for each practice, i need the following:

  1. what was the pain
  2. how they solved it
  3. why it provides continuous value to the work that they do

i need the links as well to check in case i want to study more

the practice must be based on the actual work they do otherwise, consider its bait i don't need any fancy exaggerated practices at all, simply the ACTUAL case works for the person or a micro company

once you finished listing all, go check again and validate on top of it, think about other angles and ideas to search until there's nothing comes up i'm open to WACKY ideas and angles too

my goal here is to understand the real business examples to diverge my view of LLM and how to use it more effectively

Run details and counts
Model
Sonnet 5.5
Reasoning
Medium
Operator
Anthropic
Mode
Not recorded
Search type
Not recorded
Run date
Oct 18, 2026
Pack
None
Content language
English
Length
5625 words
Links
25
Domains
8
Tables
3
Output
Report contents

Real LLM use by non-engineers and micro companies: round 3 (new cases only)

Date: 2026-10-18. Scope: new cases that do not appear in practical-llm-work-by-non-engineers-2026-10-15.md, llm-business-cases-non-engineers-2026-10-16.md or practical-llm-work-publication-check-2026-10-15.md.

Summary

I kept 17 new cases: 5 graded A, 11 graded B, 1 graded C. Two of the 17 are negative evidence (an abandoned practice and a dropped tool habit).

Most of the new cases are Japanese first-person posts on note.com. English, Chinese, Korean, French, German and Spanish searches returned almost only guides and vendor pages. I found no keepable first-person case from Taiwan, Korea, Latin America or Africa. Treat the geography as a gap in what I could reach, not as evidence that nobody does this there.

Three patterns repeat across the strongest cases.

  1. Writers who count yen, minutes and failures are more believable than writers who report "80% saved". The best cases name a price the model got wrong, a drawing it misread, or a setting that cost them five hours.
  2. The model fills blanks with plausible defaults. In three separate cases it invented a unit price, a payroll closing date and a "this month's" blood pressure value. The fix in each case was a rule, not a better model.
  3. The daily saving often comes from a plain tool that the model helped set up once (voice shortcuts, a LINE order form, a premise sheet). The model is not in the loop every day.

At least seven of the 17 authors sell prompt packs, courses, coaching or consulting. I mark this in each case. It does not make the account false, but it explains why the claims are cheerful.

How to read the evidence

  • A: first-person account fetched and read, specific work described.
  • B: third-party report of a named person, or a thin or partly promotional first-person account.
  • C: very thin.
  • "Evidence" lines come from the page. "Inference" lines are mine.
  • "(ja)" means the source is in Japanese. Summaries are mine. Quotes are short fragments translated by me.
  • URLs for note.com posts are https://note.com/<author>/n/<key>.

Cases

1. A solo seitai (manual therapy) owner pastes one "clinic premise" block, then cuts posting volume (A)

  • Who: author "PARA", runs a one-person seitai clinic in Oita, Japan. Posts dated 2026-07-30, 2026-08-01, 2026-08-05 (ja).
  • Pain: Evidence: he retyped the same background in every new chat (city, patients aged 40 to 60, "no medical claims"). Posting drafts got faster but total time did not fall. Later, more posts across five channels created a bigger problem: what goes where, and what he had already said.
  • Solution: Evidence:
    • He writes one premise block once and pastes it at the start of a chat: trade and area, target patients, common complaints, what the clinic values, how to book, first-person style, and banned words (for example "always cures").
    • After that he gives only the topic. He still adds field detail by hand ("many customers drive long hours").
    • He keeps patient names and visit history out of the chat and updates the block when prices or hours change.
    • He gave each channel one job: website blog for search, Google posts to show the clinic is open, Hot Pepper blog for pre-booking worries, note for peers, social posts short. Google posts go out on even days and the Hot Pepper blog on Fridays.
    • He cut from five post types a day to three he can check. He also uses the model to check overlap with past posts and to cut a long blog into a short Google post.
  • Money and site edits: Evidence: the company-built site cost 260,000 yen with 7,700 yen a month for upkeep, and a colour change once cost 180,000 yen. He now asks the model for small edits. Rules after breakage: keep the old code, change one thing at a time, check on PC and phone, back up before big edits.
  • Why it keeps paying: Evidence: the premise block removes setup time from every later draft, and the channel roles stop repeat content. Inference: the volume cut matters more than the premise block, because checking time, not writing time, was the limit.
  • Caveat: no traffic or revenue numbers (he says posting did not jump and revenue has not arrived). The account has 38 notes and 15 followers, and he sells a "100 prompts" paid note. I could not corroborate the clinic's identity.
  • URLs: https://note.com/agile_gibbon4423/n/nc9a41d00d99e , https://note.com/agile_gibbon4423/n/n6aca7700c9d7 , https://note.com/agile_gibbon4423/n/nfe662d5bd3de

2. A neighbourhood electric shop photographs receipts, and documents what the model could not read (A)

  • Who: Kazuhiro Kaneko, nickname "Kaneden", of Kaneden Denki (Kaneko Electric), an appliance and electrical work shop in Ota-ku, Tokyo, 27 years in the field. Series dated 2026-10-06 to 2026-10-13 (ja).
  • Pain: Evidence: after site work comes a second job: photo sorting, quotes, supplier prices, reports to property managers, invoices, tomorrow's plan. He loses track of "when did I contact them" and "where is that photo". He has about 300,000 site photos in Google Photos.
  • Solution: Evidence:
    • A ChatGPT project he calls his "secretary". After buying materials he photographs the receipt and sends it with prompts such as "which site is this for", "organise it for the purchase ledger", "this job is not yet billed".
    • He checks the read amounts and item names himself, because a photo does not update the books by itself.
    • He adds fields over time (site name, purchase date, unbilled flag) by correcting the output and trying again.
    • A second tool, Manus, handles photos and bulk documents. Google Drive holds one folder per job.
  • Failures he published: Evidence:
    • He gave Manus plan PDFs and asked it to pick the lighting fixtures and propose a swap to another maker. Symbols and quantities came back wrong. His reading: drawing symbols differ by architect, and he had handed over an unfamiliar drawing and said "read it all".
    • He tested file-name patterns for finding old photos. New photos were found; he could not tell why old ones were not, and says he will not claim what he does not know.
    • Learning from videos did not stick. Trying one real task did.
    • He settled on two tools, not four.
  • Why it keeps paying: Evidence: the receipt habit happens daily, and every correction becomes a standing field. Inference: the habit works because the check step is cheap (compare a number on a photo).
  • Caveat: no time saved is stated. "Daily use" is self-report; the series starts 2026-10-06. The company did not appear in my search for it.
  • URLs: https://note.com/kaneden/n/nbf53f4db5ab7 , https://note.com/kaneden/n/n0e925a2fd44d , https://note.com/kaneden/n/n470469cf1e5d , https://note.com/kaneden/n/n75a7d4b04bf6

3. A 30-year electrician uses the model as a "what do I usually overlook" list, and got burned on prices (A)

  • Who: Omori (author name on note), representative of Energy Futures Co., an electrical contractor in Ryugasaki, Ibaraki, 30 years in the trade. Post dated 2026-07-23 (ja).
  • Pain: Evidence: estimates leave out items that appear after submission. Site prep lists took time to build, and forgotten tools or materials happened two or three times a month. After jobs he found customer messages hard to write.
  • Solution: Evidence:
    • Estimates: he types the job (for example LED swap of 20 fixtures, 5 m ceiling, 3 wiring extensions) and asks for easily missed items, materials and usual extra work. He uses the answer as a checklist, not as an answer. Words that matter: "easily overlooked", "list format".
    • Prep: the evening before, by phone, he describes the visit and gets a list of tools, materials and checks.
    • Customer LINE after the job: he gives the work done and asks for a warm, not stiff message under 200 characters.
  • Failure he published: Evidence: the model gave a unit price for flexible conduit, he used it in an estimate, and it did not match his purchase price. Rule: the model does not know prices; check prices and law changes yourself.
  • Why it keeps paying: Evidence: each of the three tasks recurs every job. Inference: the "overlooked items" prompt is the most transferable, since it asks the model for omissions instead of answers.
  • Caveat: the title claims two extra hours; I found no measurement in the text. He promotes a paid member club. I could not corroborate the company in search. His note profile links to business social pages that I did not open.
  • URL: https://note.com/rodori_4647/n/n9bbb48039fc4

4. A one-person LLC owner has Claude clean up a year of messy books (A)

  • Who: Glenn Hopper, finance professional with about 25 years in finance, runs a single-member LLC for speaking, training, books and fractional work. Post dated 2026-02-06 (en).
  • Pain: Evidence: books never closed, personal and business cards mixed, all revenue in one "Sales of Product Income" account, and he had avoided his bookkeeper for 14 months.
  • Solution: Evidence:
    • Claude Cowork drove QuickBooks Online in the browser. It rebuilt revenue, cost and expense accounts, reclassified past entries and produced a cash-basis P&L. He says this took under an hour. Balance-sheet and depreciation work remained.
    • He exported 2025 card CSVs, had it flag likely business expenses using travel dates, training costs and known subscriptions, then used Claude for Excel to add an expense account per line. For this step he states he stayed "the human in the loop".
    • He had it structure quarterly journal entries for owner distributions.
    • 1099s he did with QuickBooks' built-in generator, no AI.
  • Why it keeps paying: Inference: the job is yearly, but the supplied context (travel dates, subscriptions) can be reused each year.
  • Caveat: he also works on enterprise AI, so he is not a typical non-engineer reader. The page does not describe reconciliation. He did not send the entries to his bookkeeper. He skipped vetting third-party QuickBooks connectors. This is a US account.
  • URL: https://glennhopper.substack.com/p/i-let-an-ai-agent-do-my-year-end

5. A German-to-English translator translates first, and asks the model for options only when stuck (A)

  • Who: Kristal Fellinger, German-to-English translator and copywriter who helps German small companies reach new customers. ATA German Language Division blog, 2026-03-06 (en).
  • Pain: Evidence: marketing text has to read as natural English, and clunky sentences cost time near deadlines.
  • Solution: Evidence:
    • She translates by hand first. She says she does not toss the source text into a tool and post-edit.
    • When a sentence sounds clunky she asks for about five options (removing identifying details when needed) and combines pieces with her own version.
    • When time is short she uses the model to gain distance from her own wording.
    • She asks it to explain idioms and puns, then verifies.
    • Test on one IT-refurbishment paragraph: with a plain "translate this" prompt ChatGPT read more naturally than Claude; with a detailed prompt she preferred one Claude phrase; none matched her own restructure of the last sentence.
  • Why it keeps paying: Evidence: it fits every job without handing over the text. Inference: "options, not output" keeps her style and her liability.
  • Caveat: employment status is not stated. The comparison covers one paragraph.
  • URL: https://ata-divisions.org/GLD/german-to-english-marketing-translations-how-ai-can-help-and-hinder-us/

6. A dairy farm that also runs cheese sales and a one-party inn moves orders into LINE (B)

  • Who: Takeshita Kosuke (author name on note), second-generation farmer at Takeshita Farm, Nakashibetsu, Hokkaido (dairy, cheese, one-party inn). Post dated 2026-10-09 (ja).
  • Pain: Evidence: staff tried Notion, Slack, LINE WORKS and new accounts, then returned to their own phones. Orders, delivery slips and shipping labels needed the same names and quantities written again and again.
  • Solution: Evidence:
    • Wholesale cheese orders arrive through a LINE official account (32 partners, not all moved yet). A Wednesday message opens orders for Friday shipping. Orders land in a spreadsheet by partner, totals feed production, and the Friday delivery slip (freee) and Yamato shipping label connect to it.
    • Cleaning staff for the inn have a second LINE account with a rich menu: schedule, checklist, report, handoff.
    • Next step, unfinished: staff speak at day's end, an AI asks follow-up questions, the staff member confirms, and the day's record and the handover note are made. He warns that AI must not turn "something I noticed" into "something went wrong".
    • He builds the tools with Claude Code, without writing code line by line.
  • Why it keeps paying: Evidence: the same weekly order cycle. Inference: the entry point (an app staff already use) matters more than the model.
  • Caveat: the voice report is not in use. I cannot tell from the text where a language model runs inside the order flow; it may be mostly ordinary automation. This is a builder case. The farm and inn exist: the MAFF Hokkaido site and a Mynavi Agri page appeared in my search results (not opened).
  • URL: https://note.com/takeshitafarm/n/n83c5f8c03ad1

7. A seitai salon owner moves from chat to folder-based work with Claude (B)

  • Who: Director Takeyama of the seitai salon "360 laboratory", interviewed by Shigeyuki Kameda of HuX, who coaches him (ja, 2026-09-25).
  • Pain: Evidence: he depends on Hot Pepper for customers and has no time or skills for other channels.
  • Solution: Evidence: he moved from chatting with ChatGPT to keeping a folder of past exchanges for Claude. Main uses are analysing Salon Board data and drafting note articles.
  • Why it keeps paying: Evidence: both tasks recur. Next wishes (not done): a CRM joining LINE and Hot Pepper customers, lapsed-customer mail, coupons, posters.
  • Caveat: the post says the Q&A is rebuilt from interview notes, not verbatim. "About 80% less work" and "CVR +0.5 points" have no stated task, period or other changes, and the author says these are still being confirmed. The interviewer is a paid coach.
  • URL: https://note.com/kamechi/n/nb1dca481087e

8. A therapist at a seitai clinic orders two opposed critics in one prompt (B)

  • Who: a physical therapist who says he works at Shinkoiwa Ekimae Seitai (employee or owner not stated). Post dated 2026-06-18 (ja).
  • Pain: Evidence: after two months of daily use the model said "good idea" to everything, which does not help with Instagram posts, flyers, clinic management or patient care.
  • Solution: Evidence: a standing instruction to answer from two views: a clinic owner aiming at a second site (field, acquisition, staff development) and an outsider executive earning 100 million yen (margin, repeatability, systems). Each answer must give good points, problems, pros, cons, alternatives, and disagreement where it exists.
  • Why it keeps paying: Evidence: he reports more criticism and says some of it is wrong but useful as sparring. Inference: the cost is one paragraph, once.
  • Caveat: no examples of decisions changed and no numbers.
  • URL: https://note.com/shinkoiwa_0301/n/n5cecea6c82e6

9. A home piano teacher plans a recital with ChatGPT (B)

  • Who: a teacher with a small home studio for nearly 20 years. Post dated 2026-02-03 (ja). The account has one note.
  • Pain: Evidence: running everything alone; a recital every 1.5 years means date, venue, notices to each parent, pieces, program order, fee, day-of flow.
  • Solution: Evidence: (1) a parent notice built from her bullets; (2) a fee suggested from venue, souvenir and headcount ("a third-party view", she decided); (3) a program order from grade, mood of piece, duets and closer; (4) day-of flow and a short MC script. Rules: no names, vague descriptions of students, she decides.
  • Why it keeps paying: weak. Inference: tasks recur only every 1.5 years, so the saving is real but infrequent.
  • Caveat: no time numbers; single post; identity not corroborated.
  • URL: https://note.com/sayapiano74/n/n91dbca1b58fe

10. A pharmacist uses the model once per drug class to build voice shortcuts for charting (B)

  • Who: a pharmacist who writes under "Pharmacist AI time-saving, Keu" and works at a pharmacy with many long-term prescriptions. Post dated 2026-05-26 (ja).
  • Pain: Evidence: many stable patients on the same lifestyle-disease drugs; the pharmacist wants assessment and plan text that is not identical for every disease, because uniform notes can fail an audit or be indefensible after an adverse event.
  • Solution: Evidence: templates are made per drug class by adverse effect (ARB: high potassium, low blood pressure, kidney, fatigue; DPP-4: low blood sugar, pancreatitis, pemphigoid, bowel change). ChatGPT organises the effects once; the author rewrites in their own words and checks anything unfamiliar. In the pharmacy's voice-input system, a short spoken code like "ARB1 circle cross" inserts the text plus a reading and adherence mark. Subjective notes stay in the patient's words.
  • Why it keeps paying: Inference: the model is used once; the daily saving is the voice shortcut.
  • Caveat: employee at a pharmacy, no time numbers. Template charting has its own audit risk.
  • URL: https://note.com/teijigaeri/n/nd8042d5c63a2

11. A part-time London tutor keeps Notion AI as a "second memory" (B)

  • Who: Sam Finnegan-Dehn, charity fundraiser by day, part-time maths and philosophy tutor for university students. MIT Technology Review, 2026-06-02 (en).
  • Pain: Evidence: he lacks time to grow his client list and struggles to turn a long-term goal into steps.
  • Solution: Evidence: Notion AI at $20 a month, chosen because his tutoring notes are in Notion. With consent it records sessions and summarises them; he adjusts teaching when a summary shows a technique is not working. It reminds him about client progress and new readings, drafts lesson notes and invoices, and posts to social media. He sets a "North Star" goal and the AI proposes steps. He does not use AI for teaching materials.
  • Why it keeps paying: Evidence: he calls it a "second memory" that links ideas across notes.
  • Caveat: no numbers; he finds the tool "clunky"; the article raises privacy concerns about the tool's access. Third-party report.
  • URL: https://www.technologyreview.com/2026/06/02/1138227/how-small-businesses-can-leverage-ai/

12. A registry-office assistant tests AI on neighbour letters and decides not to build it (B)

  • Who: Takuro Okada, 15 years as an assistant in land-registration offices, now trains such offices on AI. Post dated 2026-06-13 (ja).
  • Pain/test: Evidence: neighbour greeting letters for boundary confirmation look easy to automate. He asked ChatGPT to research and propose automation.
  • Result: Evidence: a 12,000-character answer, accurate but too heavy for a first letter (civil code article, ADR, lawsuits). A "for elderly readers" sample used childish hiragana. Given the office's steps (map, registry records, owner addresses, Word template plus Excel mail merge, post, phone follow-up) it grew into a whole "boundary meeting intake system". He felt "this is troublesome" and concluded it adds steps beside an existing template.
  • Why it matters: Evidence: his rule is that if AI adds input, reading, double management or checking, it is added work, not improvement.
  • Caveat: a consultant's trial, not a routine. Whether the office later adopted anything is unknown.
  • URL: https://note.com/takorou21/n/n3159d4d8ff6e

13. A labour-law consultant compares three models on work rules and keeps a "pre-decision" list (B)

  • Who: Kazuko Kume, labour and social-insurance consultant (sharoushi), Hiroshima, running an office she names Reiwa. Posts dated 2026-05-09 and 2026-07-05 (ja).
  • Test: Evidence: the same bare request, "make a draft of work rules", to ChatGPT (paid), Claude (free) and Gemini (free). ChatGPT asked a few questions, then wrote specific pay dates (closing on the 25th, pay on the 10th) that she never gave. Claude had the best questions and layout but also invented pay dates. Gemini wrote about three pages without questions. All three omitted a leave-of-absence rule and fell short of the ministry's model work rules (10 or more chapters, about 70 articles). Her conclusion: none is usable now; the issue is information, not the tool, and she plans follow-ups with intake sheets and the model text.
  • Routine use: Evidence: she uses ChatGPT to set out issues, options, missing checks and items to hold before a decision. She tells it not to assert, uses dummy data for personal information, and reads primary documents herself. She keeps 14 behaviour rules from these chats.
  • Why it keeps paying: Inference: decision prep recurs daily; the "missing checks" step catches blind spots.
  • Caveat: plans differ (free versus paid) and use history differs. She sells AI-implementation support. A Hiroshima office named Reiwa appears in a Money Forward partner listing (seen in search, not opened), but I did not confirm the person.
  • URLs: https://note.com/sr_reiwa/n/nda7692d9d1a3 , https://note.com/sr_reiwa/n/nfcdcd5d3de0b

14. A welding and fabrication shop owner has ChatGPT search past orders before re-ordering (B)

  • Who: owner of a metal fabrication shop in Aichi (the profile name is Lumi Spark). Post dated 2026-10-09 (ja).
  • Pain: Evidence: ordering materials means reading the drawing, picking out items and sending an order. One wrong dimension ruins material. A round plate was missing from the order.
  • Solution: Evidence: ChatGPT, connected to his Gmail and Google Drive, searched past order mails and the purchase ledger. It found earlier orders with size, processing, quantity and unit price, then noticed that a hand-written inner diameter on the drawing had been corrected with the supplier earlier. It marked the drawing in red with a candidate value. He says it is only a candidate and the current drawing revision must be checked.
  • Why it keeps paying: Inference: the value is in joining records that sit in different places, which repeats every order.
  • Caveat: one episode. The account started in July 2026 and mostly posts about a local community project. Letting a model search company mail has a privacy cost the post does not discuss.
  • URL: https://note.com/lumispark/n/n7b001224f0ec

15. A bar owner spends five hours fixing numbers that the AI-written sheet got wrong (B, builder)

  • Who: owner of a bar in Kokura, Fukuoka. Post dated 2026-09-23 (ja).
  • Pain/story: Evidence: in May 2026 he had a model write Google Sheets and Apps Script code for weekly figures. The table and the AI summary disagreed on new customers.
  • Cause: Evidence: for 27 April the source had 2 new guests and 1 new group, while the imported sheet showed 0 for both. Blank, unread and true zero values were all treated as 0. Fixes to the text and table did not repair the missing import.
  • Rules he adopted: Evidence: keep a restorable copy; pick one record with a known answer; state what may and may not be changed; decide the success test before the fix; do not treat "two outputs agree" as proof, compare with the source.
  • Caveat: this is a builder case. He sells how-to material. It is useful mainly as a warning about "blank equals zero".
  • URL: https://note.com/shogo0623/n/ne36f25f1ec4b

Negative evidence

16. A consultant's own test: AI-drafted first-touch emails halved the reply rate (B)

  • Who: Lilach Bullock, British AI and marketing consultant. Undated ("last spring"), (en).
  • Evidence: month 1, AI drafts: 62 sent, 9 replies. Month 2, AI drafts with little editing: 58 sent, 4 replies. Month 3, she wrote the opening line and used AI for the body: 54 sent, 14 replies. She now writes the opener, deletes generic lines, reads each email aloud and sends smaller batches. She still uses AI daily for subject lines, research and admin.
  • Caveat: small sample, self-reported; she sells AI consulting; her "open-to-reply" label is wrong, since it is plain reply rate.
  • URL: https://www.lilachbullock.com/why-i-stopped-letting-ai-write-my-emails/

17. A translator drops tools to train for a certification exam (C)

  • Who: Jordan Levine, Japanese-to-English translator, ATA Japanese Language Division blog, 2026-07-06 (en).
  • Evidence: he writes that habitual reliance on Trados, DeepL and ChatGPT hurt his ability to translate on his own, and he practises without them in a weekly group of three to four because the exam bans them.
  • Caveat: one sentence of evidence; unclear for which tasks or whether he still uses them for client work.
  • URL: https://www.ata-divisions.org/JLD/2026/07/06/2026_jpen_translationstudygroup/

What the validation changed

I re-read all URLs. For 17 note.com posts I re-pulled the page text and checked 3 to 5 specific claims each (all found). I re-opened the MIT, ATA, Hopper, Bullock and Levine pages with targeted questions.

  • Hopper: my first note said his bookkeeper would receive the files. He actually never contacted her. Fixed. Also, his "human in the loop" statement covers card categorisation only.
  • Fellinger: I had "never post-edits raw AI output". The page says she does not toss the source text into a tool and post-edit. "Self-employed" is not stated. Fixed.
  • Finnegan-Dehn: all five claims confirmed.
  • Bullock: figures confirmed, but no date, and she sells AI consulting. Graded B and used as negative evidence only.
  • Downgrades: Takeshita (B, builder, key AI feature unfinished); the piano teacher (B, recurrence only every 1.5 years); Okada (B, a trial); Lumi Spark (B, one episode); Levine (C).
  • Dropped after reading: 15 candidates (see table below).
  • Identity checks: Takeshita Farm and FARM VILLA taku appear in third-party search results (MAFF Hokkaido, Mynavi Agri, travel sites). Kaneden Denki, Energy Futures, Shinkoiwa Ekimae Seitai and the Oita clinic did not appear in my searches. Treat those four as unverified first-person accounts. Their note profiles show 15 to 287 followers and 38 to 151 notes, which fits ordinary small accounts.

Cases found and dropped

CandidateSourceWhy dropped
"Tsukimatsu", home-visit acupuncturist writing about eight AI mistakes in reportsnote.com/tsukimatsu_ai (ja)Account is a few days old (first post 2026-10-11, about 12 long posts since), part paywalled, and the examples are built on invented sample records. Possible bait. The failure types (a blood pressure not measured this month, a date shifted) match real cases 3 and 15, so I note them as hypotheses.
Ryokan owner learning AI for phone bookingsnote.com/great_human8287 (ja)Two inns, likely not micro, and he says the app is not yet in use.
Temple priest starting an AI projectnote.com/vivid_emu3238 (ja)First post, plans only.
"Ramen shop owner: posting time down to 5 minutes a week"note.com/raumen_ai (ja)Figures (+420 followers, +18% weekend guests) unverifiable; ends with a paid note; author profile says restaurant adviser.
Izakaya owner "Kenichi Tanaka", 3 hours to 10 minutesnote.com/arukuwa (ja)A fictional story with invented characters, written by a support service. This is the commonest bait pattern.
Part-timer who made a pork cutlet photo look better with Gemininote.com/chatoran555 (ja)Not the owner's work; the post is about a doctored food image.
Physical therapist's "valley between demo and daily use"note.com/kimu1019 (ja)Useful list of five reasons peers drop AI, but he sells setup help and the 13 to 17 automations cannot be checked. Kept as hypothesis only.
Care-business on-device model guidenote.com/chic_wren6567 (ja)A consultant's proposal with sample data, no use.
Florist "AI x flower shop" LP seriesnote.com/hanaya_ai (ja)Series begins 2026-10-07; no routine yet.
Ex-restaurant owner now selling automationnote.com/mute_mint5020 (ja)Seller of builds; trap list is for coding agents.
IT-consultant farm and writing seriesnote.com/tototo51 (ja)Runs a consulting firm; project just starting.
Bakery founder Bo&mie, podcast 2024podcast.ausha.co (fr)Page has no concrete use; audio not available.
Gangnam market training, Sud Radio artisans, Yahoo "3 owners"variousMerchants unnamed or uses described are not ChatGPT or are thin.
Senter, Arkansas farmer (DTN 2024)dtnpf.comAlready case 2 of the first note.
ATA JLD interpreter-tools summaryata-divisions.orgSecond-hand summary of a talk; only prep prompts.

Search angles tried and where I stopped

AngleResult
Japanese note.com via its search API, about 90 queries across roughly 60 trades (seitai, acupuncture, salons, nails, cafe and bar, izakaya, ramen, ryokan, guesthouse, farms, dairy, fishing, forestry, electricians, painters, mechanics, architects, surveyors, scriveners, labour consultants, pharmacy, vets, care managers, child care, after-school clubs, tutors, piano and calligraphy teachers, yoga, translators, guides, landlords, funeral, shrines and temples, handmade sellers)The main source of the 17 cases. The last three rounds (shrines, funeral, after-school care, nails, osteopaths, cram schools, handmade) returned no keepable new case.
Failure and "I stopped" wording in JapaneseGave cases 3, 12, 15, 16 and the dropped bait. Few honest "abandoned" posts exist; most failure posts end with a sales pitch.
English general search by trade (piano tuner, groomer, tutor, childminder, dance studio, property manager, church admin, beekeeper, farm stand, funeral celebrant, electrician)Guides and vendor pages only. One keep (Finnegan-Dehn, via news).
News search (Bing news RSS) for "owner says she uses ChatGPT"Surfaced MIT Technology Review and a Yahoo piece; mostly vendor launches.
Translator and interpreter association blogs (ATA divisions, NOTIS)Cases 5 and 17. A NOTIS 2023 post (one prompt session) was too old and one-off.
Taiwan: Chinese queries for minshuku, small farmers, shop owners, Vocus, Threads, PTTNothing first-person. PTT and Vocus search were not readable by my tools.
Korea: Korean queriesOnly a Gangnam market training notice (no named merchants).
France, Germany, Spanish-speaking countries, AfricaTraining and survey pages, vendor guides, one podcast blurb. A German craft-trade search returned only advisors.
Reddit trade forumsNot searched this round. The mirrors I could reach were behind a bot check, and I did not try to pass it.
Hacker News, Substack search, WordPress.com reader searchNo usable first-person non-engineer cases; WordPress results were SEO spam.

I stopped because three consecutive rounds returned only sellers, fiction, duplicates, or trades with no first-person posts. Reddit trade forums and Taiwanese platforms remain the likeliest untouched sources, but my tools could not read them this round.

Untested angles and ideas (hypotheses)

IdeaWhy it might matterHow to test
Use the model to interview a retiring craftsperson (voice, one question at a time) and turn it into a handover manualSeveral guides describe this; I found no real case.Ask one senior staff member for a 30-minute recorded interview; have the model draft a one-page procedure; ask them to mark every wrong line.
Reading bad handwriting (delivery slips, old ledgers)A US seed-farm story from a vendor claims this; I found no independent one.Photograph 20 handwritten slips; compare model transcription with the original; count wrong digits.
Using the model to word a refusal or a price negotiationIllustrator fee-floor case (earlier note) suggests it.Take five real "no" or counter-offer emails; compare staff drafts with model drafts for tone, then send the better.
Care and grief work (hospice, funeral staff)Only obituaries and generic advice surfaced.Out of scope for the company; skip unless a client asks.
A dated "traps" file per task typeCases 3, 13 and 15 show the same recurring blanks.See experiment 3 below.

What a reader can copy

  1. Write the premise once (case 1). Paste a block with audience, tone, banned claims and booking method at the start of work. Keep personal data out. Refresh it when prices change.
  2. Ask for omissions, not answers (cases 3, 13). "What do people usually overlook in this job?" and "list the missing checks" returned usable lists. Direct answers returned invented prices and invented dates.
  3. Plan for blank filling (cases 3, 13, 15). Tell the model to leave a field empty or ask, and compare any number with a source. Never let blank become zero.
  4. Limit output to what you can check (case 1). More posts created a management problem; he cut volume.
  5. Use the model once to set up something plain (cases 6, 10). A voice shortcut or a LINE form pays every day with no model call.
  6. Count the extra steps (case 12). If the AI version adds steps beside a working template, stop.
  7. Ask for disagreement on purpose (case 8). One paragraph gives you two critics.
  8. Photograph the paper (case 2). Receipts and slips are the cheapest daily input.
  9. Be careful who is talking. At least seven of the 17 authors sell something. The most useful posts were those that said what went wrong and what it cost.

Five small experiments for a 20-person Japan and Taiwan-facing consulting and web company (non-engineering staff)

Each runs two weeks with three volunteers. Stop early if it adds work.

  1. Premise sheet per client (cases 1, 13). Each volunteer writes a one-page block for two clients: audience in Japan or Taiwan, tone, banned claims, terms never to translate literally, and who approves. Paste it at the start of every draft. Measure: minutes from request to first usable draft, and edits per draft, with and without the sheet. Keep client names and personal data out.
  2. "What did we overlook" check on proposals (case 3). Before sending a quote, paste the scope and ask: "List items clients later add, and items missing from this scope." Staff keep the list as a checklist and never accept prices. Measure: how many listed items turn out to be real changes later, and how many are noise.
  3. Blank-filling trap file (cases 3, 13, 15). Keep a shared page with dated one-line entries: what the model invented (a price, a date, a legal term, a translated phrase) and the rule that follows. Add a line only after verifying the cause. Measure: how many entries each person logs, and whether repeats fall.
  4. Two-critic review (case 8). For each proposal draft, ask for two reviews: a procurement lead at a Japanese client and an outsider with no industry knowledge. Compare with the human reviewer's notes. Measure: points caught by the model that the human missed, and points that were wrong.
  5. Gap test before any automation (cases 12, 13). Pick one routine document (for example a Japanese business email template or a service agreement). Have two models draft it from a bare brief, then score against your own must-have checklist. Do this before building a prompt or flow. Measure: omissions and invented defaults per model. If the AI path adds steps compared with the current template, do not build it.

Sources not opened

Mynavi Agri (https://agri.mynavi.jp/2025_06_04_312832/) and MAFF Hokkaido (https://www.maff.go.jp/hokkaido/kushiro/douga/takeshitabokujo_cheese360.html) appeared in search results for Takeshita Farm. The Money Forward partner directory listing for Reiwa Sharoushi also appeared in results. I did not open these pages.